Structural resilience evaluation is crucial for ensuring structural safety, with structural damage detection (SDD) serving as a core component. Although convolutional neural networks (CNNs) have been proven effective in extracting damage-sensitive features for SDD, their limited receptive field and weak global information processing during feature extraction can lead to insufficient accuracy and reduced stability when dealing with field monitoring data. To tackle these issues, a multi-scale Transformer network called Transformer neural network with multi-scale acceleration feature fusion (TMAFF) is presented for enhanced SDD results. Both overlap patch embedding and dilated overlap patch embedding (DOPE) are employed to acquire feature representations for acceleration at different scales respectively, which reserve more features in acceleration signals and thereby improve the method’s robustness against noisy or incompletely measured data. DOPE employs dilated convolution to enlarge the receptive field of convolutional layers without increasing additional parameters. Additionally, an improved Transformer block, integrating cross-channel correlation-based multi-head attention (CCMA) and gated depth-wise separable convolution feedforward network mechanisms, is included for automated extraction of damage-sensitive features. The use of CCMA improves calculating efficiency and spatial correlation consideration compared to traditional self-attention-based Transformers. Furthermore, employing the GELU activation function enhances the neural network’s nonlinear modeling and complex data distribution fitting capabilities, thereby improving the accuracy and stability of SDD. Numerical simulations conducted on a five-storey rigid frame indicates that TMAFF outperforms CNN in feature extraction, providing more reliable accuracy for SDD. Validations using field monitoring data from both an experimental frame and the Yonghe Bridge demonstrate the feasibility of the proposed method for structural resilience evaluation, even with a limited number of sensors available.

中文翻译：

---

使用现场监测数据基于多尺度 Transformer 网络的自动结构弹性评估


结构弹性评估对于确保结构安全至关重要，其中结构损伤检测（SDD）是其核心组成部分。尽管卷积神经网络（CNN）已被证明在提取SDD损伤敏感特征方面是有效的，但其有限的感受野和特征提取过程中全局信息处理能力较弱，可能会导致在处理现场监测数据时精度不足和稳定性降低。为了解决这些问题，提出了一种称为具有多尺度加速特征融合的 Transformer 神经网络（TMAFF）的多尺度 Transformer 网络，以增强 SDD 结果。重叠贴片嵌入和扩张重叠贴片嵌入（DOPE）分别用于获取不同尺度下加速度的特征表示，保留了加速度信号中的更多特征，从而提高了方法对噪声或不完整测量数据的鲁棒性。 DOPE采用扩张卷积来扩大卷积层的感受野，而不增加额外的参数。此外，还包括一个改进的 Transformer 模块，集成了基于跨通道相关的多头注意力 (CCMA) 和门控深度可分离卷积前馈网络机制，用于自动提取损伤敏感特征。与传统的基于自注意力的 Transformer 相比，CCMA 的使用提高了计算效率和空间相关性考虑。此外，采用GELU激活函数增强了神经网络的非线性建模和复杂数据分布拟合能力，从而提高了SDD的准确性和稳定性。 在五层刚架上进行的数值模拟表明，TMAFF 在特征提取方面优于 CNN，为 SDD 提供了更可靠的精度。使用来自实验框架和永和大桥的现场监测数据进行的验证证明了所提出的结构弹性评估方法的可行性，即使可用的传感器数量有限。